ESSAYS ON INTERNATIONAL TRADE AND LABOR MARKET OUTCOMES. Tommaso Tempesti. Dissertation. Submitted to the Faculty of the

Transcription

1 ESSAYS ON INTERNATIONAL TRADE AND LABOR MARKET OUTCOMES By Tommaso Tempesti Dissertation Submitted to the Faculty of the Graduate School of Vanderbilt University in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY in Economics August, 2011 Nashville, Tennessee Approved: Professor Robert Driskill Professor Yanqin Fan Professor Andrea Moro Professor David Parsley Professor Joel Rodrigue

2 Copyright c2011 by Tommaso Tempesti All Rights Reserved

3 To the memory of Matteo Lanzoni. ii

4 ACKNOWLEDGMENTS I am of course indebted to the suggestions and advice received from my advisor Robert Driskill and other members of my dissertation committee, Yanqin Fan, Andrea Moro and Joel Rodrigue. I am however responsible for any remaining errors. Bob Driskill offered me another chance when I had very few. He then wisely lead me through the dissertation and job market process and also honored me of his friendship. Priceless, as they say. Yanqin Fan kindly agreed to be in my committee and gave me important suggestions as to my econometric approach. Interacting with Andrea exposed me again to the intellectual ruthlessness that I had come to miss in the courteous South. All of this would not however have been possible without the help of Joel Rodrigue. My research got a new start after I attended his seminar on empirical international trade. He supported my original dissertation idea, gave me comments on the substance of my papers while also patiently advising me on the details of building my dataset. This is all the more remarkable given that he is a young assistant professor. Many other persons deserve to be acknowledged. Eric Bond, Bob Hammond and Sangsoo Park gave me very useful comments on a draft of the second chapter. I have also benefitted from the professional advice of Jeremy Atack and Bill Collins in their capacity of Director of Graduate Studies. I had the pleasure of working first at the Center for Evaluation and Program Implementation at Peabody and then at the Department of General Pediatrics at the Vanderbilt Medical Center. These jobs not only provided me with important funding but also with the opportunity of professional growth. Special thanks go to Ana Regina Vides de Andrade, Warren Lambert and Shari Barkin for their kind support. I spent four academic semesters in McTyeire, an international language dormitory on the Vanderbilt iii

5 campus. This gave me the opportunity to live, cheaply, close to campus at a crucial time in my academic career. It was also a great experience in and of itself as I met very talented students and friends. Special thanks go to Anja Bandas and to the members of the Italian hall. I cannot express my gratefulness to my parents and my brother who never failed to support me. A special thank goes to my brother Duccio who visited me in 2008 and spent three great weeks with me. I thank my old friend Salvatore Florio for being always willing to help me. I greatly benefitted from having as friends and colleagues Mostafa Beshkar, Linda Carter, Bob Hammond, Alisina Önder, Sangsoo Park and Jisong Wu. A special thank goes to Sangsoo with whom I shared many graduate school s experiences. Alessandro, Lucia, Rita and all the other italian friends in Nashville offered me fun and relief for several years. As to the friends on the other side of the pond, the trips with Niccolò, whether in Europe or in U.S., have always been something to look forward. Finally, I thank Vanderbilt University for giving me the opportunity to get a postgraduate education of excellent quality and the city of Nashville for being my home during these years. It has been a pleasure to witness, and be part of, the recent growth of both communities. iv

6 TABLE OF CONTENTS Page ACKNOWLEDGMENTS... LIST OF TABLES LIST OF FIGURES iii vii ix Chapter I INTRODUCTION II LITERATURE REVIEW Introduction Empirical Motivation Labor Market Outcomes Trade Flows Empirical Methodology: Regressions Trade And The Labor Market Trade and the Skill-Premium The Stolper-Samuelson Model Technology, Quality Upgrading and Offshoring Trade and Residual Inequality Dynamic Effects of Trade Conclusion III OFFSHORING AND THE SKILL-PREMIUM Introduction Theory Discussion Data Description Results Wages: Individual-Level Regressions Wages: Industry-Level Regressions Employment Conclusion Data Appendix Industry Level Data Worker Level Data IV OFFSHORING AND OCCUPATIONAL SWITCH Introduction Literature Review Data and Empirical Model Worker-Level Data v

7 Industry-Level Data Empirical Model Results Conclusion Data Appendix Worker Level Data Industry Level Data V CONCLUSION BIBLIOGRAPHY vi

8 LIST OF TABLES Table Page 1 Literature Review: Summary Literature Review: Summary, continued Summary statistics Individual wage regressions by education category, with year and industry dummies Individual wage regressions by education, without year and industry dummies, U.S Individual wage regressions by education, no demographics, no year and industry dummies. U.S Industry-level wage regressions by education, no year and industry dummies. U.S Industry-level wage regressions by education, no year and industry dummies. U.S , continued Industry-level wage regressions by education category without demographics. U.S Industry-level hours of employment regr. by education, no demog., no year and industry dummies Summary statistics Switch of Occupation Regression Switch of Occupation Regression, corrected for attrition vii

9 14 Switch of Occupation Regression, continued. Attrition Model Switch of Occupation Regression on Imported Parts Switch of Occupation Regression, Attrition Model Switch of Occupation Regression on Imported Parts, corrected for attrition viii

10 LIST OF FIGURES Figure Page 1 Production Workers Employment in U.S. Manuf., Import Penetration in U.S. Manufacturing, The Skill-Premium In U.S. in the 1980s Offshoring and Skill-biased Technological Change in U.S. Manufacturing: White Workers Share of Employment in U.S. Manufacturing: Average Job Experience in U.S. Manufacturing: Female Workers Share of Employment in U.S. Manufacturing: ix

11 CHAPTER I INTRODUCTION The effects of globalization on domestic economies have been, and still are, widely debated. In this thesis I contribute to the literature on the effects of international trade on labor market outcomes. In the first chapter I review the more recent work on the subject and notice how the use of datasets that contain data on individual workers is promising for advancing the field. In the second chapter I combine micro level data on U.S. workers from the March Current Population Survey in the 1980s with macro level data on trade for the industries the workers are in. In particular, I focus on trade in intermediate inputs, that I call offshoring. Offshoring in U.S. manufacturing grew more than 25% between 1970 and 1990 (Hummels et al. (2001)). I find an effect of offshoring on the wage premium paid to educated workers with respect to less educated ones. This skill-premium is an important component of wage inequality. In the third chapter, I study if offshoring has also affected residual inequality, i.e. the wage inequality that is not explained by basic observable characteristics of the worker. In order to do this, I study if offshoring has affected the workers probability of switching their occupations. Previous studies have found that the increase in occupational switching accounts for a large portion of the increase in residual inequality. I find however that offshoring does not affect occupational switching. Taken together, my results imply that, at least for U.S. in the 1980s, offshoring increased wage inequality by increasing the skill-premium but did not affect residual wage inequality. More specifically, the first chapter surveys recent empirical works that study the effects of trade on labor market outcomes. The focus is on studies that use individual workers data. These data allow to control for the changing variation of the labor force 1

12 across industries over time. In this regard, these studies improve over previous ones that do not control for worker level variables. I first review works that are based on regressions. These works find that, at least for the U.S., trade either increases wage inequality or has no effect on it. I discuss the reasons that may explain the variety of these results. These results seem to suggest that, at least for the U.S., the increase in trade did not benefit the poorest among the workers. This is in line to what Goldberg and Pavcnik (2007) find in their review of studies on developing countries. I then review empirical works based on structural models. These models suggest instead that studies based on wage regressions neglect the long-term consequences of trade on the reallocation of resources across sectors and the efficiency gains associated with it. I finally show the usefulness of matched employer-employee datasets and discuss along the way promising avenues for future research. In recent decades many countries have experienced an increase in both international trade and the skill-premium. The association between these two phenomena has proven elusive in the early empirical literature on the subject. Indeed, the consensus among labor economists seems to be that trade has not been the main cause of such increase in the skill-premium. This view has been challenged by Feenstra and Hanson (1999) who find that offshoring sizably affects the skill-premium. In the second chapter I revisit this debate using individual workers data from the March Current Population Survey combined with industry-level trade data. This strategy improves upon the work of Feenstra and Hanson who do not control for the demographic characteristics of the labor force. I show that industry-level wage regressions overestimate the impact of offshoring on the skill-premium if the demographic characteristics of the labor force are omitted. In addition, I find that offshoring increases the relative employment of skilled workers, thus suggesting that offshoring has played an important role in the increase in the skill-premium by increasing the 2

13 economy-wide relative demand of skilled workers. Various studies have stressed the importance of an increase in economic turbulence for the understanding of the labor market. Kambourov and Manovskii (2009b) document an increase in the fraction of workers switching occupations since the early 1970s to the mid-1990s. They show how this increase is able to explain a substantial portion of the concurrent increase in residual wage inequality. They attribute the increase in occupational switching to the increased turbulence in the economy. Offshoring seems to be a possible candidate for the increase of turbulence in the labor market. In the third chapter, using data from the March Current Population Survey for the period, I study if offshoring in manufacturing is correlated with occupational switching. I find that offshoring does not increase the probability of switching occupations. The coefficient on offshoring is either non-significantly different from zero or significantly different from zero and negative. This result is robust to the use of different measures of offshoring and to controlling for attrition out of the March Current Population Survey from one year to the other. This result suggests that offshoring from U.S. to abroad has not been responsible for the increase in residual inequality. 3

14 CHAPTER II LITERATURE REVIEW Introduction Recently, the impact of international trade on labor market outcomes has received a great deal of attention in the press. 1 At the same time, the academic literature has devoted an increasing attention to the impact of international trade on labor market outcomes. In particular, many studies have combined data on individual workers with industry level data to study e.g. the impact of trade on inequality, on the wages of the unskilled workers and on unemployment. The goal of this paper is to survey the recent empirical literature in this area and to point to promising avenues for future research. Some clarifications are in order here. First, this is not a survey of the theoretical literature on trade and the labor market. I will refer to that literature when relevant but the focus here is on the empirical work, its challenges and results. Second, this is not even a survey of all the empirical literature on the subject. Indeed, I concentrate on studies that make use of individual level data on workers. 2 There is for example a large literature in the 1990s about the effect of trade on wage inequality but that literature mainly uses data on workers that is aggregated at the industry level. 3 Again, I will refer to this literature when relevant but the focus here is on empirical studies that use individual data on workers. 1 Irwin (2009) and Wolf (2004) detail the various policy debates concerning globalization. Amiti et al. (2005) document the importance of service offshoring during the 2004 U.S. presidential election. 2 A bit of historical perspective may be helpful. One of the most significant development in the recent trade literature is the use, especially since the end of the 90s, of micro level data on firms or plants. A subsequent development has been the use of micro level data on workers, which is the focus of this survey. These two approaches are now being combined in the use of employer-employee matched datasets. In section II.4.3 I review Muendler and Menezes-Filho (2007) who utilize a employer-employee matched dataset. 3 This literature is reviewed in e.g. Acemoglu (2002, Section 6.3). 4

15 Even this narrow focus does not allow me to review all the more recent papers in the literature given their increasing number. Rather my survey is only selective with the intent to delineate some of the main features and challenges of the work in this area. In the study of the impact of trade on the labor market, the usefulness of individual data on workers can be illustrated in several ways. First, we know that a worker s wage is affected by her observable individual characteristics. Studies that rely only on aggregate wages at the industry level and do not control for the individual characteristics of the workers within each industry omit an important determinant of the level of wages in each industry. 4 If the characteristics of the workers are correlated with the trade regressor of interest, the empirical results will be biased. Including industry fixed effects in the regression (or differencing over time within an industry) alleviates this problem only in part because trade may also involve a changing composition of the labor force within an industry over time. Moreover, even if the characteristics of workers were not correlated with the regressors of interest, their inclusion in the wage regression will likely decrease the standard errors on the estimates on the relevant regressors. A worker s wage is also affected by her unobservable individual characteristics. If workers sort across industries based on these unobservable characteristics, then the estimates of the parameters of interest may be biased. Using longitudinal individual worker s level data also allows to control for worker fixed effects. It also also allows the study of other relevant outcomes. For example, one can observe the movement of workers across industries, across occupations and across labor market status (e.g. employed in the formal sector, employed in the informal sector, unemployed). 4 The same is true of studies that use data on workers that is aggregated at the firm or plant level without controlling for the individual characteristics of the workers within each firm or plant. 5

16 Goldberg and Pavcnik (2007) review the literature on the distributional effects of globalization in developing countries. This survey is different from theirs in the following ways. First, the focus of this survey is on studies that utilize individual worker s data. While their paper also reviews some of such studies, this paper updates theirs focusing on more recent developments in this area. Second, many studies that utilize individual worker s data refer to the U.S. experience. Because of this focus I review many papers that study the U.S. labor market and not the developing countries one. Finally, I use the Mincerian equation to organize the exposition of the material and, to the best of my knowledge, my paper is the first to do that. But some distinctions in the literature are standard (e.g. the distinction between inter-sectoral vs. intra-sectoral reallocation of resources) and Goldberg and Pavcnik (2007) articulate them in an excellent way: I therefore follow their work in some regards and I document when I do so. Another related work is Crinò (2009). He focuses on the labor market effect of offshoring of services, offshoring of manufacturing and foreign direct investment. He reviews only few studies that use individual worker s data and none that uses a structural model. He also does not survey the more recent literature on the labor market effects of trade in final goods while I do not survey the literature on the international movement of capital. In this sense, our two surveys are then complementary. Section II.2 reviews the stylized facts for the labor market and for trade flows that motivate the works in this area. Section II.3 reviews the methodologies that are usually utilized in the literature, with a special focus on Mincerian wage regressions. Section II.4.1 reviews the results on the effects of trade on the wage premium paid to skilled workers with respect to unskilled workers. Section II.4.2 reviews the results on the effects of trade on other components of wage inequality. The studies reviewed in these sections use wage regressions and find that, at least for U.S., trade either increases wage inequality or has no 6

17 effect on it. The reasons that may explain the variety of these results are discussed in section II.3. From these results one could be tempted to conclude that the increase in trade did not benefit the poorest among the U.S. workers. This would be in line to what Goldberg and Pavcnik (2007) find in their review of studies on developing countries. However, structural models suggest that studies based on wage regressions neglect the long-term consequences of trade on the reallocation of resources across sectors and the efficiency gains associated with it. These studies are reviewed in section II.4.3. This section also reviews a study that uses matched employer-employee data and shows the advantages of using such data. Table 1 contains an overview of the results. Section II.5 concludes. 7

18 Table 1. Literature Review: Summary Study Country Methodology Dataset Results Period N Feenstra and Hanson (1999) U.S. Two-stage regression Industry Panel Offshoring accounts for at least % of increase in skill-premium Lovely and Richardson (2000) U.S. OLS with individual Worker Panel (PSID) Trade with newly industrialized countries fixed effects 6,477 does not increase the industry premium to skilled relatively to unskilled workers Kosteas (2008) U.S. OLS with individual Worker Panel (NLSY) Wage semi-elasticity w.r.t. imports share fixed effects N/A from low-wage countries is -.6 Ebenstein et al. (2009) U.S. OLS Worker Cross-section Wage semi-elasticity w.r.t. import share (CPS MORG) at occupation level is ,505,724 8 Liu and Trefler (2008) U.S. First Differences Worker Cross-Section Offshoring of services does not (Matched March CPS) significantly affect wages 37,550 Attanasio et al. (2004) Colombia First Differences Worker Cross-Section Wage semi-elasticity w.r.t. tariffs (Household Survey) is Source: compilation of the author.

19 Table 2. Literature Review: Summary, continued Study Country Methodology Dataset Results Period N Artuc et al. (2010) U.S. Structural Estimation Worker Cross-Section Workers in the import-competing sector (March CPS) sector may benefit from a lower tariff N/A because of higher option value of moving to another sector Artuc (2009) U.S. Structural Estimation Worker Panel (NLSY) Middle-aged workers in the import-competing sector are hurt the most by a lower tariff because of industry-specific human capital 9 Muendler et al. (2007) Brazil Logit with Matched Employer- After trade liberalization, job separations individual fixed effects Employee, 1% random are higher, and job accessions are lower, sample from population in comparative-advantage sectors and at exporters. Source: compilation of the author.

20 Empirical Motivation Labor Market Outcomes A large literature has documented the changes in the labor market in recent decades, in both developing and developed countries. Figure 1 graphs the employment of production workers in U.S. manufacturing during the period. Even though the U.S. population increased during this period, the number of production workers employed in U.S. manufacturing declined by almost 1,400, Using data from the Current Population Survey (CPS hereafter), Ebenstein et al. (2009) estimate a decrease in total employment in U.S. manufacturing from 22 millions in 1979 to 17 millions in 2002, with a rapid decrease in the more recent years. A possible explanation for this decline is that the price of foreign manufacturing goods has decreased relatively to the price of domestic manufacturing goods and so the U.S. consumers have substituted away from domestic manufacturing. Several studies have also documented the changes in the wage structure in U.S. and in other countries. In this regard it is useful to introduce the Mincerian wage equation that has been used extensively in the labor literature. A popular version of this equation is the following: ln(w s )=α + β 1 Race s + β 2 Gender s + β 3 Exp s + β 4 Exp 2 s + β 5 Educ s + s (II.1) where s is a worker, ln is the natural logarithm, w s her wage, Race s a set of race dummies, Gender s a gender dummy, Exp s a proxy for work experience, Educ s a measure of education such as years of schooling or a set of dummies for degrees completed, and s an error assumed 5 Data are from the NBER productivity database. I graph the data since 1972 which is the earliest available year for the trade data used in Figure 2. The number of total workers (production and non-production) employed in U.S. manufacturing declined by around 700,000 during this period. 10

21 to be random. If we measure inequality of wages with the variance, then this equation allows to decompose wage inequality in two components: the variance of the predicted wage and the variance of the estimated residual. 6 The first term is also called between-inequality because it captures that portion of wage inequality that depends on e.g. men earning on average more than women. The second term is also called within-inequality because it captures that portion of wage inequality that does not depend on the demographic characteristics of workers. In other words, there is inequality of wages even among workers who have the same demographic characteristics. This second term is also called residual inequality, which is the term I will use in this review. A remarkable change in the wage structure has been the increase in the wage premium paid to educated workers. This is also called skill-premium under the assumption that education is a good proxy of the skill of the worker. 7 In terms of equation (II.1), this means an increase in β 5 over time. 8 Autor et al. (2008, Figure 2, p.303) show that, since 1979 to 2005, the average wage paid to a worker with a college degree grew 20% more than the average wage paid to a worker with only a high-school degree. Given that the relative employment of U.S. skilled workers has also increased, this suggests that the relative demand for skilled workers has increased. 9 Similar patterns for relative employment and relative wages of skilled workers have also been detected in many developing countries, 6 More specifically, the Mincerian equation allows to decompose the inequality of the log of wages. The logarithm is useful for several reasons. First, it reduces the impact of very large incomes on the estimates. Second, the distribution of the log of wages is fairly normal which is useful when using ordinary least squares. Third, the use of logs facilitates the interpretation of the coefficients that become either elasticities (if the regressor is also in logs) or semi-elasticities. In what follows, when I talk about wage I actually mean the log of wage. 7 Some studies proxy unskilled workers with production workers and skilled workers with non-production workers. 8 Sometimes equation II.1 is run in different years allowing the coefficients to change over time. 9 This point has been made several times in the literature. See for example Katz and Autor (1999), Acemoglu (2002) and more recently Crinò (2009, p.203). 11

22 especially since the 1980s. 10 The increase in the skill-premium has also contributed to the increase in overall wage inequality. In U.S. the wage of workers at the 90th percentile has increased around 25% more than the wage of workers at the 10th percentile (see Autor et al. (2008, Figure 2, p.303)). 11 In Section II.4.1 I discuss the relationship between various forms of international trade and the increase in the relative demand for skill. The coefficients on the other variables in equation II.1 have also changed over time, at least for U.S, and contributed to the evolution of wage inequality over time. 12 There are only few works that discuss how international trade has affected these coefficients and so this is possibly an area for future research. Even if the vector of β coefficients does not change, wage inequality may still increase because of an increase in residual inequality. The evidence on residual inequality is mixed. For the U.S. Katz and Autor (1999) at first estimated that the increase in residual inequality could explain up to 2/3 of the increase in overall inequality. Kambourov and Manovskii (2009b, Figure 1, p.736) use the PSID and find that the contribution of the residual inequality to the increase in overall inequality is very large. Lemieux (2006) argues instead that around 3/4 of the increase in residual inequality between 1973 and 2003 disappears when controlling for compositional effects. 13 Bertola and Ichino (1995, Figure 2) document that residual inequality was fairly stable in Great Britain, France and Italy during the 70s and the 80s. Attanasio et al. (2004) document a sizable increase, albeit non- 10 See the review in Goldberg and Pavcnik (2007, Table 1, p.48). 11 Obviously, income inequality need not translate in consumption inequality. Early work found evidence that consumption inequality also markedly increased in U.S. in the 1980s (Cutler and Katz (1992)). However, recent work that uses the Consumer Expenditure Survey has reconsidered this result showing that consumption inequality has increased much more moderately than income inequality (Krueger and Perri (2006)). 12 See Katz and Autor (1999). 13 In other words, the residual in equation (II.1) is heteroskedastic: the variance of the residual depends on the specific combination of observable characteristics of the worker (e.g. older workers tend to have more dispersed wages). Lemieux (2006) shows that a large part of the increase in the residual inequality can be explained by the increase in the size of the combinations that have a higher variance of the residual (e.g. the workforce getting older). 12

23 monotonic, in residual inequality between 1984 and 1998 for Colombia. In section II.4.2 I discuss the relationship between trade and residual inequality. Trade Flows The recent decades have also witnessed a remarkable increase in trade across countries. In Figure 2 I graph the import penetration in U.S. manufacturing for the period. Import penetration was 7% in 1972 but it increased to 22% in In its 1998 annual report the WTO documents that, worldwide, merchandise exports grew by 6 per cent in real terms from 1948 to 1997, compared to an annual average output growth of 3.7 per cent (p.33). It also reports that in developed countries, openness measured by the ratio of trade to GDP increased from 16.6 to 24.1 per cent between 1985 and In developing countries this indicator rose from 22.8 per cent to 38.0 per cent over the same period (p.33). Finally they also report how the composition of trade has been changing, moving away from agriculture to manufacturing, with the more recent rise in trade in services. 15 The academic literature has also focused on the specific forms in which trade can take place. Many studies have documented how it is not just trade in final goods that has increased but also trade in intermediate inputs, a phenomenon also referred to as offshoring. Crinò (2009, p.198) documents that offshoring in manufacturing, measured as the share of 14 The original data is from Bernard et al. (2006). This data is available at Peter Schott s website. Import penetration by SIC 87 industry is defined as imports/(shipments-exports+imports). Data is available for all years for only 386 industries out of 459 SIC 87 industries. I take the simple average of import penetration across industries, within a year. I also computed the average of import penetration across industries, within a year, using the share of employment in an industry as weight. The data on employment by SIC 87 industry is from the NBER productivity database. The weighted import penetration was 6% in 1972 and 16% in The pattern of growth over time is similar for both measures of import penetration. 15 The same document indeed notices that agricultural exports accounted for almost 47 per cent of total merchandise exports in 1950, and their share had dropped to 12 per cent by Manufactures, by contrast, accounted for 38 per cent of exports in This share increased to 77 per cent by 1996 and that services trade in OECD countries increased at almost twice the rate of merchandise trade between 1980 and 1995 (p.34). 13

24 imported intermediate inputs in total non-energy input purchases, has increased in the United States from 5.1% in 1972 to 18.1% in Hummels et al. (2001, Figure 2) shows how this phenomenon is common to many industrialized countries. Crinò (2009, p.199) also documents the rapid increase of service offshoring since the 90s in many industrialized countries. 16 Empirical Methodology: Regressions Many studies use a regression approach to estimate the impact of trade on wages. A common strategy consists in adding trade-related variables to the Mincerian equation (II.1): ln(w sit )=βx st + γt it + λ i + µ t + η sit (II.2) where w sit is the wage of worker s at time t in industry i, β and γ are a vector of coefficients, X st is a vector of worker s variables, T it is a vector of trade-related measures for time t in industry i, λ i are industry dummies, µ t are time dummies and η sit is the error term. Sometimes the regression comprises also the interaction between T it and some element of X st such as education. This formulation allows for the case in which the dataset is a repeated cross-section of workers and the case in which the dataset is a panel of individual workers. One challenge of regression (II.2) is that the identification of γ is obtained by the variation of trade across industries. 17 Because of this fact, one would like to have a fairly large number of industries in the sample. On the other hand, the more disaggregated the 16 The focus of this paper is on trade flows. However, in recent decades the movement of capital and labor across countries has also increased. See Crinò (2009, Figure 1, p.199) for the growth in world FDI outflows and Borjas et al. (1997, Figure 1, p.5) for the increase in immigration to U.S.. 17 More precisely, given the presence of industry and time fixed effects, the identification of γ is obtained by the variation of changes in trade, over time, across industries. 14

25 industries are, the more likely it is that a shock to a certain industry will have general equilibrium effects on the other industries. Many studies find however only a small reallocation of workers across industries after a trade shock and so in the literature many have felt that such concerns can be ignored. 18 Another concern in regression (II.2) is measurement error in the regressors of interest. For the trade variable, T it, some studies use a quantity variable, such as the value of imports, whereas others use a price variable such as a tariff. 19 The advantage of using the latter is that usually models have predictions for the relationship between e.g. the price of imports and the price of factors. The disadvantages however are that accurate price data are hard to obtain and that such data may also be misleading. For example, a change in tariff does not capture any change in the world price that the importing economy faces. But if a country is large, then a change in its tariff may have an effect on world prices. Even if the country is small, changes in tariffs may be correlated with changes in world prices, such as an increase in productivity abroad, that may affect the estimate of the coefficients on the price variables. Moreover, there may be barriers to trade other than tariffs so that a change in tariffs need not translate directly into change in domestic prices. So, even if the choice of the volume of trade as regressor is usually dictated by data availability, the quantity approach, though not ideal, is not necessarily unjustified. Whatever trade variable one uses, (II.2) may still be plagued by endogeneity. Time dummies control for macroeconomic phenomena that affect all sectors and industry dummies control for time-invariant differences across sectors. However, trade may be correlated with some other variable that varies over time at the industry level. This could be productivity, output or technological change. For this reason many studies include other industry-level 18 See the references in section II Obviously, the value of imports is not a pure quantity variable because it depends on the price of imports as well. 15

26 variables in regression (II.2) in the hope to reduce this form of endogeneity. In this regard, however, it is especially concerning that it is hard to have a good proxy for technological change. Even if T it is exogenous or can be instrumented for, there could also still be endogeneity for the individual-level variables. For example, skill is usually proxied with some measure of education. But if more able workers self-select into schooling, then the coefficient on education will not capture the effect of skill on wages. This may be a problem for studies that want to estimate the impact of T it on wages across different levels of education. Panel data on workers allows to estimate (II.2) adding individual fixed effects that account for time-invariant individual ability. The challenges of studies that use panel data on workers are discussed in Section II when reviewing the work by Lovely and Richardson (2000) and Kosteas (2008). A regression such as (II.2) relies on the assumption that the trade shock is unexpected. Some studies use as trade shock a specific episode in time. 20 If one can argue that such shock was unanticipated by the economic actors, then one can compare outcomes before and after the shock to identify the parameters of interest. However, usually the parameters in (II.2) are identified using the variation in trade across industries over some period of time. The assumptions required for exogeneity are that the economic actors not only do not anticipate the first trade shock (assuming the first trade shock hits the economy in the first year of the sample) but also that they do not adjust their decisions after they have been hit by such shock (or have observed others being hit by such shocks). These concerns have motivated the use of structural model that take explicitly into account the 20 For example, Verhoogen (2008) use as quasi-experiment the sudden, and arguably unanticipated, exchange-rate depreciation of 1994 in Mexico. 16

27 forward-looking aspect of economic decisions. 21 I examine some of these models in section II.4.3. Even if η sit is truly exogenous, for hypothesis testing it is important to have unbiased estimates of the standard errors. In (II.2) the assumption of independence of the errors across observations will probably be violated. First, it is possible that the residuals of workers who are in the same industry at the same time will be correlated. A strategy to deal with this is to first regress wages on individual characteristics, take an average of the residuals within industry-year cells and then regress these average residuals on industry level variables, industry dummies and time dummies. 22 One could be instead be tempted to cluster the standard errors at the industry-time level. However, it is also possible that the residuals of workers who are in the same industry at different points in time will still be correlated. Bertrand et al. (2004) show that this serial correlation may severely underestimate the standard errors and therefore lead to a high probability of Type I error. Using simulations, they determine that clustering standard errors at the industry level, rather than at the industry-year level, will deal satisfactorily with this problem, as long as the number of clusters is large. 23 Therefore many studies that use regression (II.2) cluster the standard errors at the industry level. Trade And The Labor Market 21 This problem is, at least in theory, distinct from the possibility of feedback effects across industries mentioned above. Even if there is no feedback across industries, a worker may react to the shock by changing jobs within its industry or moving out of the labor force. On the other hand, even if a shock to an industry is unexpected, it may be transmitted to other industries as well. 22 In this way one reduces the dataset to a panel of industries. In this case it is possible to also use first-differences to estimate the relationship between wage and trade. 23 Bertrand et al. (2004) suggest to have around 50 clusters but it is below 20 clusters that serial correlation impacts standard errors severely. 17

28 Trade and the Skill-Premium The Stolper-Samuelson Model The Stolper-Samuelson model has been used to explain the increase in wage inequality between skilled and unskilled workers. 24 In the simplest version of this model an economy produces two goods with two factors, skilled and unskilled labor, and exchanges them with the rest of the world. One good, call it A, is intensive in unskilled labor relative to the other good, call it B, which is then intensive in skilled labor. An economy that is abundant in skilled labor, such as the U.S., will export the good that is intensive in skilled labor and import the good that is intensive in unskilled labor. As trade becomes less costly, U.S. will export more of good B and import more of good A. So, in U.S. resources will have to reallocate from the production of good B to the production of good A. Given that B is intensive in skilled labor, this reallocation will increase the relative demand for skilled workers. Given that in this model the relative supply of labor is fixed, the increase in the relative demand will increase wage inequality between skilled and unskilled (assuming, as usual, that even before the shock skilled workers already were earning more than the unskilled workers). The model implies that the opposite process is happening in the country that U.S. trades with: production is reallocated from good B to good A and the relative demand for unskilled workers decreases wage inequality between skilled and unskilled workers. Finally, another implication of the model is that, as the relative wage of skilled workers increase, each sector substitutes away from skilled labor so that the skilled workers share employment is lower in both sectors. 24 See e.g. the discussion in Goldberg and Pavcnik (2007, p.58). 18

29 This model has however found scant support in the data. Many studies have indeed found that the industry employment shares are pretty constant over time: the reallocation of resources across sectors that is dictated by the Stolper-Samuelson model does not seem to take place. 25 Moreover, wage inequality between skilled and unskilled workers has increased in both developed countries, which are abundant in skilled workers, and in developing countries, which are abundant in unskilled workers. This contradicts the implication of the model that inequality should go down in developing countries. Finally, several studies have documented that the skilled workers share of employment has increased in all sectors, contradicting another implication of the model. 26 Technology, Quality Upgrading and Offshoring Because of the empirical problems of the Stolper-Samuelson model, other models have been proposed. These models emphasize how trade may affect the skill-premium within sectors. Some studies focus on the link between trade and skilled-biased technological change (SBTC hereafter), which is a usual candidate for the explanation of the increase in skillpremium. These studies show how SBTC can be an endogenous response to competition from abroad. For example, in Thoenig and Verdier (2003) s model, domestic firms, when faced with competition from abroad, engage in more innovation, which is usually a skillintensive activity. If for each sector one measures competition from abroad with the volume of sectoral imports, then an implication of this theory is that higher imports will increase the relative demand for skilled workers within that sector. Other models focus on the complementarity between capital and skill: this may affect the skill-premium if trade allows 25 See the references in Goldberg and Pavcnik (2007, p.59) and Cosar (2010, fn.4). Section II.4.3 discusses studies that allow for intersectoral reallocation of labor. 26 See Berman et al. (1994) for the U.S. and Attanasio et al. (2004) for Colombia. 19

30 to import capital goods at a cheaper price. 27 Alternatively, trade liberalization may be responsible for quality upgrading within one sector, i.e. the shift of the product mix toward higher quality varieties. If these higher quality varieties require a more skilled labor force, this will also increase the relative demand for skill (see Verhoogen (2008)). Feenstra and Hanson (1996a) emphasize instead the role of trade in intermediate goods within an industry. Suppose that, in order to produce a final good, several inputs are needed and that we can order these inputs as to their skill-intensity. A country that is skill-abundant will tend to produce the skill-intensive inputs and to import the inputs that are intensive in unskilled labor. As developing countries become more productive they will export to U.S. more of the inputs that are intensive in unskilled labor in U.S.. For this reason, an increase in productivity in the developing countries will shift resources away, in U.S., from the inputs intensive in unskilled labor to skill-intensive ones. Again, this fact will push up the relative demand of skill within each industry. Lovely and Richardson (2000) is one of the first works to utilize individual level workers data to study the effect of trade on the skill-premium in the U.S.. They use, for the period, the Panel Study of Income Dynamics (PSID hereafter), which follows a panel of U.S. workers over time. They focus on the effect of imports and exports, disaggregated by source country and use two estimation approaches. In the first one they regress individual wages on individual workers characteristics (including education), a set of industry dummies and a set of industry dummies interacted with years of schooling of the workers. The first set of industry dummies can be considered as the industry wage premia to labor who does not have any schooling. The second set of industry dummies can 27 See the review in Acemoglu (2002, p.27) and Goldberg and Pavcnik (2007, Section 5.1.3). Even if these models are related to the the ones that emphasize the importance of SBTC, they remain distinct. SBTC theories focus on the advent of new technologies, especially computer, and their impact on the demand for skilled labor. Capital-skill complementarity theories do not focus only on computers but rather on any kind of equipment. On this distinction, see also Acemoglu (2002, fn.24). 20

31 be considered as the premium to skill in the various industries. Average wages for skilled and unskilled workers can vary across industries for many reasons such as compensating differentials, sorting of workers across industries based on unobserved ability and rents. 28 Lovely and Richardson (2000) emphasize the first interpretation but their first approach cannot exclude the other possibilities. They then regress these estimated industry premia (to pure labor and to skill) on various trade measures, other industry level measures and year dummies. They find that trade with the newly industrialized countries tends to increase the industry premium to skilled workers relative to the industry premium to the unskilled workers. They interpret this as evidence that trade with newly industrialized countries increases the relative demand for skilled workers. If one interprets trade with newly industrialized countries as trade in capital goods or trade in varieties that have lower quality with respect to U.S. or trade in intermediate inputs, then this result is compatible with the models outlined above. An advantage of the PSID is that it allows to control for individual fixed effects: if workers with different unobservable productivities sort into different industries, then the omission of these fixed effects will bias the estimate of the effects of trade on the skill premium. In their second approach Lovely and Richardson (2000) run a regression similar to II.2 but also controlling for individual fixed effects. As other regressors they use individual workers characteristics, a set of industry dummies and their interaction with years of schooling, trade and other industry level measures and their interaction with schooling. 28 In the case of compensating differentials, identically productive workers may receive different wages because of differences in relevant characteristics across industries (e.g. safety, amenities etc.). In the case of unobserved ability, the industry wage premia are due to the heterogeneity of workers ability across industries. This heterogeneity is sometimes considered to be ex-ante and so time-invariant and therefore it is modeled with the use of individual fixed effects in the wage equation. Finally, as discussed in section II.4.2, rents may arise both because of a lack of intersectoral workers mobility or because of imperfect competition in the labor market. The lack of intersectoral mobility may be due to the presence of industry-specific human capital: workers may be identical ex-ante but they are different ex-post, due to the industry they end up in. Imperfect competition, due for example to unionization, is compatible with workers being identical in their productivity. 21

32 They again find that trade with newly industrialized countries increases the skill-premium but this result is not robust to the inclusion of year dummies. This is important given that Lovely and Richardson (2000) do not control for skilled-biased technological change (SBTC), which is thought to have changed during this period. This result also suggests that controlling for individual fixed effects may be important when studying the effects of trade on the skill-premium. Kosteas (2008) uses instead the National Longitudinal Survey of Youth (NLSY hereafter) to estimate the impact of imports from low-wage countries on wages in the period. The NLSY is a panel dataset that contains observations on individuals in U.S. who were aged in He finds that, for the period, the import share is correlated with a decrease in the wage of blue collar workers and does not affect the wage of white collar workers. 29 For the period, he finds no correlation between the import share and wages of either class of workers. He controls also for outsourcing using a measure of imports of parts by industry but he does not include the interaction between outsourcing and the white collar dummy. The NLSY, as the PSID, is a longitudinal dataset and so it allows to control for individual fixed effects in the estimation of the impact of trade on wages. When these effects are included, the effect of imports on the wage of blue-collar workers increases in absolute value (it is a negative number) and remain significant; moreover, the coefficient on the interaction between the white-collar dummy and imports becomes smaller and less significant (Kosteas (2008, Table 3(a), p. 268)). 30 Therefore, the inclusion of individual fixed effects seem to be important. 29 Because the source of the trade measure changes in 1989, Kosteas (2008) has to divide the analysis in two periods. 30 To be more specific, he finds that the semi-elasticity of blue collars wage with respect to the imports share from low-wage countries is -.6 and significant at the 1% level. The coefficient on the interaction between the white-collar dummy and imports is 0.1 and significant only at the 10% level. 22